Consider compiling SIMD-from-SIMD initialisation directly to a shuffle

[huonw]

#![crate_type = "lib"]
#![feature(tuple_indexing)]

use std::simd::f32x4;

pub fn foo(x: f32x4) -> f32x4 {
    f32x4(x.0, x.2, x.3, x.1)
}

becomes, with no optimisations:

define <4 x float> @_ZN3foo20h2254f602671f886ceaaE(<4 x float>) unnamed_addr #0 {
entry-block:
  %sret_slot = alloca <4 x float>
  %x = alloca <4 x float>
  store <4 x float> %0, <4 x float>* %x
  %1 = getelementptr inbounds <4 x float>* %sret_slot, i32 0, i32 0
  %2 = getelementptr inbounds <4 x float>* %x, i32 0, i32 0
  %3 = load float* %2
  store float %3, float* %1
  %4 = getelementptr inbounds <4 x float>* %sret_slot, i32 0, i32 1
  %5 = getelementptr inbounds <4 x float>* %x, i32 0, i32 2
  %6 = load float* %5
  store float %6, float* %4
  %7 = getelementptr inbounds <4 x float>* %sret_slot, i32 0, i32 2
  %8 = getelementptr inbounds <4 x float>* %x, i32 0, i32 3
  %9 = load float* %8
  store float %9, float* %7
  %10 = getelementptr inbounds <4 x float>* %sret_slot, i32 0, i32 3
  %11 = getelementptr inbounds <4 x float>* %x, i32 0, i32 1
  %12 = load float* %11
  store float %12, float* %10
  %13 = load <4 x float>* %sret_slot
  ret <4 x float> %13
}

with optimisations it becomes

define <4 x float> @_ZN3foo20h2254f602671f886ceaaE(<4 x float>) unnamed_addr #0 {
entry-block:
  %sret_slot.12.vec.insert = shufflevector <4 x float> %0, <4 x float> undef, <4 x i32> <i32 0, i32 2, i32 3, i32 1>
  ret <4 x float> %sret_slot.12.vec.insert
}

We could detect when a SIMD vector is being created directly from elements of another (pair of*) SIMD vector(s) and convert it directly into the appropriate shuffle instruction. This will save allocas and LLVM doing work, and probably guarantees it more than LLVM currently does. This should even work for vectors of different lengths, as long as the elements are the same.

(This is just a bug since it's an implementation detail.)

*shufflevector actually takes two operands, so f32x2(x.0, y.0, x.1, y.1) can also directly become a shuffle.

[eddyb]

If we would stop using allocas in that case, it would be much easier for LLVM to combine extractelement and insertelement into shufflevector.
I have somewhat of a plan for that, and @luqmana has shown interest in helping, but I can't promise anything concrete (I've wanted to do it for more than half an year now, with #12424 and #13004 as my first attempts).

[steveklabnik]

Two things:

1. simd has been moved out, and @huonw, you've been doing a bunch of work on stuff, so I don't know how much of a bug this still is
2. the unoptimized version seems to remove one getelementptr now:

define <4 x float> @_ZN3foo20h88e8b14ab8916c70faaE(<4 x float>) unnamed_addr #0 {
entry-block:
  %dropflag_hint_7 = alloca i8
  %x = alloca <4 x float>
  store i8 61, i8* %dropflag_hint_7
  store <4 x float> %0, <4 x float>* %x, align 16
  %1 = getelementptr inbounds <4 x float>, <4 x float>* %x, i32 0, i32 0
  %2 = load float, float* %1, align 4
  %3 = insertelement <4 x float> undef, float %2, i64 0
  %4 = getelementptr inbounds <4 x float>, <4 x float>* %x, i32 0, i32 2
  %5 = load float, float* %4, align 4
  %6 = insertelement <4 x float> %3, float %5, i64 1
  %7 = getelementptr inbounds <4 x float>, <4 x float>* %x, i32 0, i32 3
  %8 = load float, float* %7, align 4
  %9 = insertelement <4 x float> %6, float %8, i64 2
  %10 = getelementptr inbounds <4 x float>, <4 x float>* %x, i32 0, i32 1
  %11 = load float, float* %10, align 4
  %12 = insertelement <4 x float> %9, float %11, i64 3
  ret <4 x float> %12
}

updated code:

#![crate_type = "lib"]
#![feature(core_simd)]

use std::simd::f32x4;

pub fn foo(x: f32x4) -> f32x4 {
        f32x4(x.0, x.2, x.3, x.1)
}

[huonw]

simd has been moved out, and @huonw, you've been doing a bunch of work on stuff, so I don't know how much of a bug this still is

This is a code-generation thing, so external SIMD would still benefit.

the unoptimized version seems to remove one getelementptr now:

There's still 4 that could theoretically be removed.

(Thanks for prompting for updates!)

[brson]

Closing wishlist.